Should You Buy an RTX 5090 or Use AWS for LLMs?
With the NVIDIA RTX 5090 having been on the market for about six months, the debate between building a powerful local machine versus renting cloud compute for Large Language Models (LLMs) is more relevant than ever. For individuals, families, or small teams diving into AI, the choice between the upfront cost of a top-tier GPU and the pay-as-you-go flexibility of a service like Amazon Web Services (AWS) is a critical one.
Let’s break down the real-world performance and costs to see which option makes more sense.
🔍 Performance Comparison
The RTX 5090, built on the Blackwell architecture, represents a significant leap in consumer-grade power. We’ll compare it to the AWS g5.2xlarge instance, which features the popular NVIDIA A10G GPU, a direct competitor in the 24 GB VRAM class.
| Feature | NVIDIA RTX 5090 (Local) | AWS g5.2xlarge (Cloud) |
|---|---|---|
| GPU Model | RTX 5090 | 1× NVIDIA A10G |
| GPU Architecture | Blackwell | Ampere |
| GPU Memory | 32 GB GDDR7 | 24 GB GDDR6 |
| GPU Memory Bandwidth | ~1,792 GB/s | ~600 GB/s |
| AI Performance (Approx.) | ~3,350 AI TOPS | ~280 INT8 TOPS |
| System vCPUs | Depends on your PC’s CPU | 8 vCPUs (AMD EPYC) |
| Typical Street Price | ~$2,200 (was $1,999 MSRP) | $0 (no upfront cost) |
| AWS On-Demand Cost | N/A | ~$1.21/hour (US East) |
| Ideal Usage | High-end single user, full control | Scalable, multi-user, flexible usage |
The RTX 5090 is a powerhouse, boasting next-gen GDDR7 memory, significantly higher bandwidth, and superior AI compute capabilities. The A10G is an enterprise-grade workhorse but is a generation behind in architecture and raw speed.
💸 Cost Comparison: 2-Year Ownership vs. Cloud Usage
Let’s calculate the total cost of ownership over two years, assuming fairly heavy, continuous use to stress-test the economics.
| Cost Factor | NVIDIA RTX 5090 (Local Build) | AWS g5.2xlarge (Cloud) |
|---|---|---|
| Initial Hardware Cost | $2,200 (GPU only) | $0 |
| Supporting PC Hardware* | ~$1,200 (CPU, motherboard, PSU, RAM) | Included in cloud pricing |
| Electricity Cost (2 yrs) | ~575W × 24h × 365d × 2 × $0.15/kWh ≈ $1,510 | Included in hourly price |
| AWS Usage Cost (2 yrs) | N/A | $1.21/hr × 24 × 365 × 2 ≈ $21,199 |
| Resale Value after 2 yrs | ~$1,100 (50% of original GPU cost) | N/A |
| Total Cost (Net) | $2,200 + $1,200 + $1,510 - $1,100 = $3,810 | $21,199 |
* A top-tier GPU requires a robust system. This includes a compatible motherboard, a 1000W+ PSU, CPU, and RAM.
Even accounting for a full PC build and worst-case electricity costs, the local RTX 5090 setup is over 5 times cheaper than running a comparable AWS instance 24/7 for two years.
👨👩👧👦 Usage Scenario: Family or Team Using LLMs
What if you’re not a lone user? How do these options stack up for a family or a small team of four people all experimenting with LLMs?
| Scenario | NVIDIA RTX 5090 (Local Server) | AWS g5.2xlarge (Cloud) |
|---|---|---|
| Simultaneous Users | 1-2 users effectively; 4 would require job queuing | Scales easily; 4 users can have 4 separate instances |
| Performance for 4 Users | Significant slowdowns, users must take turns for heavy tasks. Requires technical setup (e.g., JupyterHub) | Each user gets dedicated performance with no interference |
| Cost for 4 Users (2 yrs) | $3,810 total → $952 per user | $21,199 per instance → $84,796 total (if all run 24/7) |
| Convenience | Fully offline, private, no network latency. Requires self-maintenance and setup | Accessible anywhere with internet. AWS manages all hardware and software |
| Flexibility | Fixed capacity; scaling requires buying more hardware | Instantly scalable; spin up or shut down instances on demand |
While a local server is overwhelmingly cheaper, it introduces challenges in resource management for multiple simultaneous users. The cloud’s primary advantage is its seamless scalability and ease of access for a distributed team.
⚠️ The AWS Caveat: Reserved Instances
The AWS cost above assumes On-Demand pricing. If you commit to a 1-year or 3-year term with an AWS Savings Plan or Reserved Instance, you can reduce hourly costs by 40-60%. This lowers the 2-year cost to roughly $8,500 - $12,700, but it remains significantly more expensive than local hardware for continuous heavy use.
📋 Summary
- RTX 5090 (Local): The undisputed champion for cost-effectiveness under heavy, consistent use. You get superior performance per dollar, own the asset, and retain full control and privacy. The trade-off is the upfront cost and the need for self-maintenance.
- AWS Cloud Instances: The winner for flexibility, scalability, and convenience. Ideal for sporadic workloads, distributed teams, or projects requiring varying compute power. You pay a premium for the luxury of not managing hardware.
🚀 Final Recommendation
If you are an individual, a student, or part of a small, co-located team using LLMs regularly, buying an RTX 5090 is the superior financial and performance choice. The initial investment is paid back relatively quickly compared to cloud costs, and you end up with a more powerful machine.
If your team is distributed, your workload is unpredictable, or you need to scale up and down on demand, start with AWS. The high hourly cost is a fee for unparalleled flexibility. You avoid massive capital expenditure and only pay for what you use — perfect for businesses and freelancers with fluctuating needs.
Thanks for reading! Feel free to share your thoughts or your own setup comparisons on GitHub or X (Twitter).
